A Parametric Recognition and Design Method and System for Jewelry Structures Integrating Computer Vision

Abstract

This invention provides a parametric design method and system for jewelry structures that integrates computer vision, relating to the field of jewelry design technology. The method includes acquiring jewelry images and performing semantic segmentation to obtain component contour and texture data; extracting contact points and normal vectors based on boundary positions to determine topological connection data; converting contours and textures into size and style parameters and establishing associated constraints; using size parameters to drive geometric generation; controlling spatial alignment through assembly angles; and iteratively adjusting associated constraints to obtain a three-dimensional digital jewelry. This achieves parametric design of jewelry.

Description

Technical Field

[0001] This invention relates to the field of jewelry design technology, and in particular to a method and system for parametric recognition and design of jewelry structures that integrates computer vision. Background Technology

[0002] In the field of jewelry design, traditional parametric modeling methods primarily rely on designers manually operating computer-aided design software to complete the design of jewelry structures by drawing two-dimensional sketches or directly constructing three-dimensional geometry. This process typically requires designers to repeatedly adjust the dimensions, styles, and assembly relationships of jewelry components to match different creative needs or client customization requirements. Furthermore, some automated design systems employ template-based parametric generation technology, generating jewelry models through predefined structural libraries and dimensional variables; however, updating these templates depends on manual input and modification.

[0003] Because jewelry structures often contain multiple interconnected components with various connection methods and assembly relationships at different angles, designers need to measure or estimate the geometric constraints and assembly angles between components one by one. This makes the design process highly dependent on accumulated experience, making it difficult to quickly respond to market demands for personalized jewelry. Furthermore, when converting physical jewelry into digital models, existing scanning or reconstruction technologies often only output static triangular meshes, lacking explicit expression of component topological connections and editable parameter sets. This data format cannot directly support subsequent size adjustments, style replacements, or structural optimizations; designers must manually redefine parameter relationships, increasing repetitive work and the risk of errors.

[0004] Furthermore, existing methods struggle to automatically extract the connection types and assembly angles between jewelry components from raw image data. This typically requires manual annotation or auxiliary measurement tools, resulting in low automation of the design process. Even with point cloud data obtained from a 3D scanner, manual analysis of the contact surface normals and connection forms between components is still necessary to construct a parameter-driven assembly model. This reliance on manual intervention limits design efficiency and makes the conversion from physical objects to editable digital models time-consuming and laborious. Summary of the Invention

[0005] This invention provides a parametric recognition and design method and system for jewelry structures that integrates computer vision, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a parametric recognition and design method for jewelry structures that integrates computer vision, comprising: Acquire image data of the jewelry to be identified; Semantic segmentation is performed on the image data to identify the components and boundaries of the jewelry to be identified, and the contour data and texture data of each component are obtained. Based on the boundary positional relationship of each component, the set of contact points and contact surface normal vectors between the components are extracted to determine the connection type and assembly angle, thus obtaining topological connection data; Contour data is converted into a set of size parameters, texture data is converted into a set of style parameters, and association constraints between the set of size parameters and the set of style parameters are established based on the connection type in the topological connection data to obtain parameterized representation data. The geometry of each component is generated by using the set of size parameters in the parameterized expression data, the spatial arrangement of each component is controlled by the assembly angle in the topological connection data, and the set of size parameters is iteratively adjusted by the associated constraints to generate a three-dimensional digital ornament. The three-dimensional digital ornament carries the parametric expression data and the topological connection data, and outputs the three-dimensional digital ornament to realize the parametric design of the ornament.

[0007] In one optional embodiment, semantic segmentation is performed on the image data to identify the components and boundaries of the ornament to be identified, obtaining the contour data and texture data of each component, including: Multi-scale feature extraction is performed on image data. Spatial structure features and color distribution features of the image are extracted at different scale levels. The spatial structure features and color distribution features at each scale level are integrated through feature fusion operation to obtain multi-scale fused features. A pixel-level semantic prediction map is constructed based on multi-scale fusion features. A region growing operation is performed on the semantic prediction map to aggregate pixels with the same semantic category and continuous spatial location into candidate component regions. The region boundaries are extracted for each candidate component region, and the pixel coordinate sequence on the region boundary is used as the initial expression of the boundary contour. Curve fitting is performed on the initial representation of the boundary contour. By calculating the curvature change of the fitted curve, corner features and curve segment features on the boundary are identified. The corner features and curve segment features are organized into a segmented contour representation, which serves as the contour data of each component. Texture feature statistics are performed on the pixels in each candidate component region, the gray-level co-occurrence matrix of the pixels in the region is calculated, and texture directionality index and texture roughness index are extracted from the gray-level co-occurrence matrix. The texture directionality index and texture roughness index are used as the texture data of each component.

[0008] In one optional embodiment, based on the boundary positional relationships of each component, the set of contact points and the normal vector of the contact surface between the components are extracted to determine the connection type and assembly angle, resulting in topological connection data including: A directed distance field is constructed from the contour data of each component. The boundary normal vector distribution is obtained by solving the gradient field of the directed distance field. The directed distance fields of adjacent components are superimposed, and the set of contact points is extracted from the zero iso-surface after the directed distance fields are superimposed. A topological connectivity analysis is performed on the set of contact points, and the topologically connected contact points are divided into contact clusters. A local coordinate system is constructed for each contact cluster, and the variance distribution of the contact cluster along each coordinate axis in the local coordinate system is calculated by principal component analysis to obtain the morphological tensor of the contact cluster. Eigenvalue decomposition is performed on the morphological tensor to obtain the morphological principal axis directions and the corresponding morphological contribution. The connection type is determined according to the distribution pattern of the morphological contribution. When the morphological contribution is concentrated in a single morphological principal axis direction, it is determined to be an embedded connection. When the morphological contribution is distributed in two orthogonal morphological principal axis directions, it is determined to be an inlaid connection. When the morphological contribution is evenly distributed in each morphological principal axis direction, it is determined to be a wrapping connection. The boundary normal vector distribution of each contact cluster is spherically fitted, and the extreme direction of the fitted sphere is extracted as the contact surface normal vector. The assembly angle is obtained by calculating the rotation transformation matrix between the contact surface normal vector and the geometric center of the component. Topological connection data containing connection type, contact point set, contact surface normal vector and assembly angle is constructed.

[0009] In one optional embodiment, the contour data is converted into a size parameter set, the texture data is converted into a style parameter set, and the association constraints between the size parameter set and the style parameter set are established according to the connection type in the topological connection data, resulting in parameterized representation data including: Skeleton extraction is performed on the contour data. The skeleton structure of the component is obtained by calculating the central axis transformation of the contour. Skeleton nodes and skeleton branches are marked on the skeleton structure. The geodesic distance between skeleton nodes is used as a length-type dimension parameter, and the local contour width at the skeleton node is used as a cross-section-type dimension parameter, thus forming a set of dimension parameters that control the shape of the component. The gray-level co-occurrence matrix in the texture data is transformed in the frequency domain to extract the dominant frequency component and phase distribution features of the texture. By mapping the dominant frequency component to the texture periodicity parameter and the phase distribution feature to the texture arrangement parameter, a set of style parameters for controlling the appearance of the component is formed. Based on the connection type in the topology connection data, the assembly dependency relationship between connected components is identified. Based on the assembly dependency relationship, the association constraint is determined, and the parameterized expression data is obtained.

[0010] In one optional embodiment, the assembly dependencies between connected components are identified based on the connection types in the topology connection data. Based on the assembly dependencies, the associated constraints are determined, and the parameterized expression data is obtained, including: When the connection type is embedded connection, extract the cross-sectional dimension parameters of the embedded part and the embedded part in the contact area, and establish a geometric fit constraint that the cross-sectional dimension parameters of the embedded part are smaller than the corresponding cross-sectional dimension parameters of the embedded part. When the connection type is tessellation, extract the texture periodicity parameters of the components on both sides of the tessellation interface, and establish an interface transition constraint that maintains the continuity of the texture periodicity parameters on both sides of the interface. When the connection type is a wrapping connection, extract the length class size parameters of the wrapping component and the wrapped component, and establish a wrapping coverage constraint where the length class size parameter of the wrapping component is greater than the corresponding length class size parameter of the wrapped component. By using geometric fit constraints, interface transition constraints, and wrapping constraints as associated constraints, parameterized representation data is obtained.

[0011] In one optional embodiment, the geometry of each component is generated using the set of dimensional parameters in the parameterized expression data, the spatial arrangement of each component is controlled by the assembly angles in the topological connection data, and the set of dimensional parameters is iteratively adjusted using associative constraints to generate a three-dimensional digital ornament, including: Construct the center curves of each component based on the skeleton nodes in the size parameter set, calculate the curvature tensor of the center curve at each skeleton node, determine the placement posture of the cross-sectional profile based on the principal direction of the curvature tensor, perform non-uniform scaling in the principal direction based on the cross-sectional size parameters, and generate the solid geometry of each component by sweeping the deformed cross-sectional profile along the center curve. Boundary curves of the contact area are extracted from the topological connection data. Surface patches matching the boundary curves are identified on the solid geometry surfaces of each component. The normal vector field distribution of the surface patches is calculated. The assembly reference direction is obtained by weighted averaging of the normal vector field. The assembly angle is decomposed into a rotation component around the assembly reference direction and a translation component perpendicular to the assembly reference direction. Rotation transformation and translation transformation are performed in sequence to realize the spatial configuration of each component. After spatial configuration, a distance field is established between connected components based on the geometric fit constraints in the associated constraints. The adjustment direction of the size parameter set is determined by solving the gradient direction of the distance field. The adjustment range of the size parameter set is determined based on the magnitude of the distance field. The size parameter set is updated and the solid geometry is regenerated. This process is iterated until the distance field converges, generating a three-dimensional digital ornament.

[0012] In one optional embodiment, generating the solid geometry of each component by sweeping the deformed cross-sectional profile along the center curve includes: The center curve is parameterized by arc length to obtain a parameterized center curve. A sweep sampling point sequence is obtained by uniformly sampling at arc length intervals on the parameterized center curve. The tangent vector of the parameterized center curve at each sweep sampling point in the sweep sampling point sequence is calculated. The tangent vector is used as the sweep direction of the cross section profile. A cross section placement plane perpendicular to the tangent vector is established at each sweep sampling point. The deformed cross section profile is projected onto the cross section placement plane to obtain the cross section instance at each sweep sampling point. Calculate the torsion angle of the parameterized center curve between adjacent sweep sampling points in the sweep sampling point sequence. When the torsion angle exceeds the preset torsion threshold, insert a compensation sampling point between adjacent sweep sampling points. Calculate the torsion angle difference between the compensation sampling point and the adjacent sweep sampling point. Perform rotation compensation on the cross-sectional instance at the compensation sampling point based on the torsion angle difference. Extract the contour control points of each cross-section instance, construct a connecting surface between the corresponding contour control points of adjacent cross-section instances, obtain surface mesh patches by triangulating the connecting surface, topologically stitch all surface mesh patches to form a closed surface, fill the closed surface with voxels, and generate the solid geometry of each component.

[0013] A second aspect of this invention provides a parametric recognition and design system for jewelry structures that integrates computer vision, comprising: An image acquisition unit is used to acquire image data of the jewelry to be identified. The semantic segmentation unit is used to perform semantic segmentation on image data, identify each component and its boundary in the jewelry to be identified, and obtain the contour data and texture data of each component. The topology extraction unit is used to extract the set of contact points and contact surface normals between components based on the boundary positional relationship of each component, determine the connection type and assembly angle, and obtain topology connection data. The parameter transformation unit is used to convert contour data into a size parameter set and texture data into a style parameter set. It establishes the association constraints between the size parameter set and the style parameter set according to the connection type in the topological connection data to obtain parameterized expression data. The three-dimensional generation unit is used to drive the geometric generation of each component with the set of size parameters in the parameterized expression data, control the spatial arrangement of each component with the assembly angle in the topological connection data, and iteratively adjust the set of size parameters with associated constraints to generate a three-dimensional digital ornament. The result output unit is used to carry the parameterized expression data and the topological connection data of the three-dimensional digital ornament, and output the three-dimensional digital ornament to realize the parameterized design of the ornament.

[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] In this embodiment of the invention, contour, texture, and topological connection data are directly extracted from the jewelry image, completely replacing the traditional manual surveying and modeling process; semantic segmentation technology accurately identifies the boundaries and contact relationships of each component, automatically generating parameterized expressions for size, style, and assembly angles, reducing subjective errors in the design process to within the recognition accuracy range of computer vision; an iterative adjustment mechanism based on association constraints ensures that the size parameter set and style parameter set always maintain logical consistency, avoiding structural conflicts caused by parameter contradictions, and ensuring that the generated 3D digital jewelry fully conforms to the original design intent in terms of geometric shape and assembly relationships; parameterized table DaData decomposes jewelry into independently modifiable dimensions, styles, and topological variables. Designers only need to adjust the associated parameters to generate multiple variant designs without repeating image acquisition or underlying modeling. The topological connection data records the contact surface normals and connection types between components, giving the assembly process geometric determinism. By using the assembly angle as a control variable and combining it with the mathematical relationship of the contact point set, the fit and mechanical rationality between the components are guaranteed in principle. When outputting 3D digital jewelry, complete parametric expression data and topological connection data are carried simultaneously, providing a directly callable structured information source for subsequent manufacturing, simulation, or customized modifications. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a parametric recognition design method for jewelry structures that integrates computer vision; Figure 2 A flowchart illustrating the logic for extracting topology connection data. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the parametric recognition and design method for jewelry structures that integrates computer vision, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: Parametric recognition and design methods for jewelry structures that integrate computer vision include: Acquire image data of the jewelry to be identified; Semantic segmentation is performed on the image data to identify the components and boundaries of the jewelry to be identified, and the contour data and texture data of each component are obtained. Based on the boundary positional relationship of each component, the set of contact points and contact surface normal vectors between the components are extracted to determine the connection type and assembly angle, thus obtaining topological connection data; Contour data is converted into a set of size parameters, texture data is converted into a set of style parameters, and association constraints between the set of size parameters and the set of style parameters are established based on the connection type in the topological connection data to obtain parameterized representation data. The geometry of each component is generated by using the set of size parameters in the parameterized expression data, the spatial arrangement of each component is controlled by the assembly angle in the topological connection data, and the set of size parameters is iteratively adjusted by the associated constraints to generate a three-dimensional digital ornament. The three-dimensional digital ornament carries the parametric expression data and the topological connection data, and outputs the three-dimensional digital ornament to realize the parametric design of the ornament.

[0021] In one optional embodiment, semantic segmentation is performed on the image data to identify the components and boundaries of the ornament to be identified, obtaining the contour data and texture data of each component, including: Multi-scale feature extraction is performed on image data. Spatial structure features and color distribution features of the image are extracted at different scale levels. The spatial structure features and color distribution features at each scale level are integrated through feature fusion operation to obtain multi-scale fused features. A pixel-level semantic prediction map is constructed based on multi-scale fusion features. A region growing operation is performed on the semantic prediction map to aggregate pixels with the same semantic category and continuous spatial location into candidate component regions. The region boundaries are extracted for each candidate component region, and the pixel coordinate sequence on the region boundary is used as the initial expression of the boundary contour. Curve fitting is performed on the initial representation of the boundary contour. By calculating the curvature change of the fitted curve, corner features and curve segment features on the boundary are identified. The corner features and curve segment features are organized into a segmented contour representation, which serves as the contour data of each component. Texture feature statistics are performed on the pixels in each candidate component region, the gray-level co-occurrence matrix of the pixels in the region is calculated, and texture directionality index and texture roughness index are extracted from the gray-level co-occurrence matrix. The texture directionality index and texture roughness index are used as the texture data of each component.

[0022] In one specific implementation, when extracting multi-scale features from image data, a combination of convolutional layers with different receptive field sizes is used to process the input image in parallel at multiple resolution levels. Shallower layers preserve high-resolution spatial details of the image, capturing fine structures such as the edge direction and small engravings of the ornament; deeper layers expand the receptive field through downsampling to obtain semantic information about the overall shape and wide-range color distribution of the ornament. For spatial structure features, the gradient response maps and edge detection results of each level are extracted to characterize the directionality and continuity of the ornament's outline; for color distribution features, the hue, saturation, and brightness components of pixels are statistically analyzed at each level to form a multi-dimensional color histogram descriptor. The feature fusion operation upsamples and aligns the feature maps of different levels to a uniform resolution, then concatenates and weights them along the channel dimension, allowing the positional accuracy of the shallow layers and the semantic discriminative power of the deep layers to complement each other in the same feature representation, ultimately outputting a multi-scale fused feature map containing both spatial structure and color distribution information.

[0023] Based on a multi-scale fusion feature map, semantic category prediction is performed on each pixel to generate a pixel-level semantic prediction map. In the semantic prediction map, each pixel is assigned a corresponding component category label, such as common jewelry component categories like the main stone area, prong setting area, band area, and chain link area. A region growing operation is then performed on the semantic prediction map: using the pixel with the highest semantic category confidence as the seed point, it expands to its four or eight neighbors, gradually incorporating pixels with the same semantic category label and spatially continuous location into the same candidate component region until category inconsistency or a sharp drop in confidence occurs at the region boundary. For isolated pixel clusters with an area below a preset threshold, they are merged into the surrounding largest adjacent region to eliminate fragmented segmentation results caused by noise. After the region growing operation, the entire image is divided into several spatially continuous and semantically consistent candidate component regions, each corresponding to a component of the jewelry to be identified.

[0024] When extracting the boundary of each candidate component region, the edges are traced pixel by pixel, and the coordinates of the edge pixels are arranged sequentially to form a closed pixel coordinate sequence, which serves as the initial representation of the boundary contour. This coordinate sequence is jagged due to pixel discretization, and directly using it for subsequent geometric analysis would introduce significant errors; therefore, curve fitting is required. During fitting, piecewise polynomials or parametric spline curves are used to approximate the coordinate sequence. The control parameters of each curve segment are determined by minimizing the deviation between the fitted curve and the original coordinate points. After fitting, the curvature value is calculated point by point along the fitted curve. Locations where the curvature value is significantly greater than the surrounding points within a local range are identified as corner features, indicating a sharp change in direction in the contour. Continuous segments where the curvature value remains stable within a certain range are identified as curve segment features, indicating that the contour extends with a smooth arc. The corner features and curve segment features are arranged alternately according to their order on the contour, organized into a segmented contour representation. Each segment records its starting and ending corner coordinates, arc length, and curvature distribution parameters, serving as the contour data output for each component. This segmented representation method preserves the geometric precision of the jewelry's outline while compressing the continuous pixel coordinate sequence into a structured parameter description, facilitating the extraction and association of subsequent size parameters.

[0025] When performing texture feature statistics on pixels within each candidate component region, the grayscale values ​​of pixels within the region are quantized to a preset grayscale level, and then a grayscale co-occurrence matrix (GLCM) is constructed. The GLCM describes the frequency distribution of two pixel grayscale values ​​occurring simultaneously under specific orientations and spacing conditions, reflecting the spatial correlation between pixel grayscale values ​​within the region. To comprehensively characterize the directionality of the texture, the GLCM is calculated in four directions: 0°, 45°, 90°, and 135°, forming a multi-directional texture description set. When extracting texture directionality indices from the GLCM, the contrast and correlation values ​​of the matrices in different directions are compared. Directions with larger contrast differences correspond to the main directions of texture energy concentration, thereby quantifying the directional texture features produced by processes such as engraving, brushing, and embossing on the surface of the jewelry. The texture roughness index is measured by calculating the angular second moment and entropy value of the GLCM: a larger angular second moment indicates concentrated grayscale distribution and uniform, delicate texture; a larger entropy value indicates dispersed grayscale distribution and rough, complex texture. The texture directionality index and texture roughness index of each candidate component region in four directions are summarized to form the texture feature vector of the component, which is used as the texture data output of each component.

[0026] Texture data and contour data together constitute a complete description of each component. Contour data uses piecewise geometric parameters to characterize the shape boundaries of the component, while texture data uses statistical feature vectors to characterize the material and process features of the component's surface. The acquisition processes of these two types of data are independent but originate from the same candidate component region, ensuring the consistency of the correspondence between the size parameter set and the style parameter set in the component during the subsequent parametric expression stage. For jewelry containing multiple similar components (such as multiple stones or multiple chain links), contour data and texture data are extracted independently for each candidate component region, forming independent parametric description instances. This enables the parametric design system to perform batch management and differentiated adjustments for similar components.

[0027] In the calculation of the gray-level co-occurrence matrix, let the row and column gray-level indices of the matrix be respectively... and The direction angle is The spacing is Then the elements in the gray-level co-occurrence matrix Indicates direction ,spacing Under the condition, the gray value is With grayscale value The normalized frequency of simultaneous occurrence of pixel pairs. The second moment of angle in texture roughness. Depend on The entropy value was calculated. Depend on Calculated. The higher the value, the more uniform the texture; The larger the value, the more complex the texture; together, they constitute the quantitative basis for the texture roughness index.

[0028] like Figure 2 The diagram shown illustrates the logical flowchart for extracting topology connection data.

[0029] In one optional embodiment, based on the boundary positional relationships of each component, the set of contact points and the normal vector of the contact surface between the components are extracted to determine the connection type and assembly angle, resulting in topological connection data including: A directed distance field is constructed from the contour data of each component. The boundary normal vector distribution is obtained by solving the gradient field of the directed distance field. The directed distance fields of adjacent components are superimposed, and the set of contact points is extracted from the zero iso-surface after the directed distance fields are superimposed. A topological connectivity analysis is performed on the set of contact points, and the topologically connected contact points are divided into contact clusters. A local coordinate system is constructed for each contact cluster, and the variance distribution of the contact cluster along each coordinate axis in the local coordinate system is calculated by principal component analysis to obtain the morphological tensor of the contact cluster. Eigenvalue decomposition is performed on the morphological tensor to obtain the morphological principal axis directions and the corresponding morphological contribution. The connection type is determined according to the distribution pattern of the morphological contribution. When the morphological contribution is concentrated in a single morphological principal axis direction, it is determined to be an embedded connection. When the morphological contribution is distributed in two orthogonal morphological principal axis directions, it is determined to be an inlaid connection. When the morphological contribution is evenly distributed in each morphological principal axis direction, it is determined to be a wrapping connection. The boundary normal vector distribution of each contact cluster is spherically fitted, and the extreme direction of the fitted sphere is extracted as the contact surface normal vector. The assembly angle is obtained by calculating the rotation transformation matrix between the contact surface normal vector and the geometric center of the component. Topological connection data containing connection type, contact point set, contact surface normal vector and assembly angle is constructed.

[0030] In one specific implementation, after acquiring the contour data of each component, a directed distance field is constructed for the contour boundary of each component. A directed distance field is a method of encoding the distance from each point in space to the nearest boundary and its sign (negative inside, positive outside) into a scalar field. For each component in the ornament, its contour data is used as input, and the directed distance value is calculated pixel-by-pixel using a fast traversal method to obtain the directed distance field of that component. ,in For component indexes. Find the gradient field The gradient direction is the direction of the normal vector at the boundary of the component, thus obtaining the complete boundary normal vector distribution of each component.

[0031] For two adjacent components and The directed range fields are superimposed, and the superposition result is: The superimposed zero isosurface (i.e. The set of points (in a geometric sense) corresponds to the region where the boundaries of two components are in contact or closest to each other. On this zero isosurface, an isosurface extraction algorithm (such as a 2D version of Marching Squares) is used to progressively traverse the image grid, collecting the pixel coordinates that satisfy the zero isosurface condition as the set of contact points. This process can accurately locate the spatial position where two components physically contact or structurally connect, avoiding boundary drift problems caused by image segmentation errors.

[0032] Obtain the set of contact points Next, topological connectivity analysis is performed. Based on the spatial proximity between contact points, a density-based connectivity criterion is adopted to classify spatially adjacent points (those with an Euclidean distance less than a set threshold) as adjacent. Contact points are grouped into the same connected component, and each connected component constitutes a contact cluster. ,in This is the index for the contact cluster. For each contact cluster... Calculate the centroid of all contact point coordinates contained therein. A local coordinate system is constructed with the centroid as the origin. Within this local coordinate system, the coordinate components of each contact point within the contact cluster along the three coordinate axes are calculated, and the variances in each direction are calculated. The variances of the three directions are then combined. morphological tensor matrix Morphological tensor Essentially, it reflects the spatial distribution pattern of the contact clusters, that is, whether the contact area is distributed in a linear, planar, or volumetric manner.

[0033] For morphological tensors Eigenvalue decomposition yields three eigenvalues. And its corresponding eigenvector, which represents the morphological principal axis direction of the contact cluster. The magnitude of the eigenvalue reflects the extent of the contact cluster's extension along the corresponding principal axis direction, and is called the morphological contribution. Based on the distribution pattern of the morphological contribution, the connection type is determined according to the following rules: When Much larger and When the morphological contribution is concentrated along a single morphological axis, the contact area is linearly distributed, indicating an embedded connection. Typical examples include prong settings and pin settings. The contact area where the gemstone is embedded in the metal prong groove exhibits this linear concentrated distribution. and All significantly greater than When the morphological contribution is distributed along two orthogonal principal morphological axes, the contact area exhibits a planar distribution, indicating a setting-like connection. Typical examples include bezel setting and channel setting, where the gemstone's side largely adheres to the metal channel wall, resulting in this planar distribution. ... When the three factors (shape contribution, gemstone, and form) are nearly equal (i.e., the shape contribution is evenly distributed along the principal axes of each shape), the contact area exhibits a volumetric distribution, indicating a wrap-around connection. Typical examples include full bezel settings or sleeve structures. This type of even distribution occurs when metal materials wrap around a gemstone from multiple directions. In the above criteria, the quantitative standards for "significantly greater than" and "nearly equal" can be achieved by setting a threshold for the shape contribution ratio. For example, when... When it is identified as a concentrated distribution, when It is considered to be a balanced distribution, among which and The preset discrimination threshold.

[0034] After determining the connection type, it is necessary to further extract the contact surface normal vector and assembly angle. For each contact cluster... Collect the boundary normal vectors at each contact point inside (from the aforementioned gradient field). These normal vectors are treated as a set of points on a unit sphere, and a spherical least squares fitting method is used to fit them to a sphere. The purpose of spherical fitting is to find the pole that best represents the overall normal vector direction of the contact area, i.e., the pole direction of the fitted sphere. The pole direction is determined by the principal axis of the fitted sphere, and this pole direction is extracted as the normal vector of the contact surface. Its physical meaning is the overall orientation of the contact surface, reflecting the relative positional relationship between the two components in the contact area.

[0035] After obtaining the normal vector of the contact surface Next, the assembly angles are calculated. Two adjacent component parts are extracted separately. and Geometric center coordinates and Calculate the direction vector of the line connecting the two. The assembly angle is calculated by normalizing the contact surface normal vector. With direction vector Rotation transformation matrix between To express, Describes the Rotate to The spatial rotation operation required for alignment can be directly solved using the Rodrigues rotation formula from the cross product and dot product of the two vectors. Rotation transformation matrix. It is stored in a compact axis-angle format, that is, expressed by two quantities: the direction of rotation axis and the rotation angle, which makes it easy to directly call in the subsequent 3D modeling stage to control the spatial configuration of the parts.

[0036] Finally, the connection types, contact point sets, contact surface normals, and assembly angles of each contact cluster are summarized to construct complete topology connection data. The topology connection data is organized in a graph structure, where each node corresponds to a component, and each edge corresponds to a contact cluster between adjacent components. The edge's attribute fields store the connection type label, a list of contact point set coordinates, and the contact surface normal. and rotation transformation matrix This structured topological connection data not only fully records the geometric relationships and assembly constraints between the various components of the jewelry, but also provides clear connection semantics for the establishment of subsequent parametric representation data, enabling different connection types to correspond to different constraint generation rules, thereby supporting the precise parametric design of the jewelry structure.

[0037] In one optional embodiment, the contour data is converted into a size parameter set, the texture data is converted into a style parameter set, and the association constraints between the size parameter set and the style parameter set are established according to the connection type in the topological connection data, resulting in parameterized representation data including: Skeleton extraction is performed on the contour data. The skeleton structure of the component is obtained by calculating the central axis transformation of the contour. Skeleton nodes and skeleton branches are marked on the skeleton structure. The geodesic distance between skeleton nodes is used as a length-type dimension parameter, and the local contour width at the skeleton node is used as a cross-section-type dimension parameter, thus forming a set of dimension parameters that control the shape of the component. The gray-level co-occurrence matrix in the texture data is transformed in the frequency domain to extract the dominant frequency component and phase distribution features of the texture. By mapping the dominant frequency component to the texture periodicity parameter and the phase distribution feature to the texture arrangement parameter, a set of style parameters for controlling the appearance of the component is formed. Based on the connection type in the topology connection data, the assembly dependency relationship between connected components is identified. Based on the assembly dependency relationship, the association constraint is determined, and the parameterized expression data is obtained.

[0038] In one specific implementation, when performing skeleton extraction on the contour data, a median transformation method is used to map the two-dimensional contour region of each component to its topological skeleton. The core idea of ​​median transformation is: for each pixel within the contour region, calculate the radius of its maximum inscribed circle to the contour boundary; the trajectory of the centers of all local maximum inscribed circles constitutes the skeleton structure of that component. The skeleton structure contains several branching points and endpoints; these feature points are defined as skeleton nodes, and the connected paths between skeleton nodes are defined as skeleton branches. Skeleton node identification is achieved by statistically analyzing the connectivity of each skeleton pixel's neighborhood: pixels with a connectivity greater than 2 are branching skeleton nodes, and pixels with a connectivity equal to 1 are endpoint skeleton nodes.

[0039] After the skeleton structure is determined, the geodesic distance between skeleton nodes is used as a length-type dimensional parameter. The geodesic distance is calculated by accumulating Euclidean distances segment by segment along the skeleton path, rather than directly taking the straight-line distance between nodes. This accurately reflects the actual extension length of the component under bending or folding conditions. Let the... The first skeleton node and the first The geodesic distance between the skeleton nodes is ,but It is obtained by accumulating the Euclidean distances between all adjacent pixels along the corresponding skeletal branch. For slender components such as the ring segment of a ring or the chain segment of a necklace, It directly represents its axial unfolding length; for the main body area of ​​the pendant, multiple skeletal branches... The overall outline scale is described by combination.

[0040] Cross-sectional dimensional parameters are obtained by measuring the local profile width at skeleton nodes. At each skeleton node, the measurement extends laterally along a direction perpendicular to the skeleton tangent until it reaches the profile boundary. The distance between the two boundary points is the cross-sectional width at that node, denoted as . ,in This is the index of the skeleton node. This reflects the cross-sectional dimensions of the component at that location, such as the thickness of the ring face, the diameter of the chain, and the width of the prongs. It includes all the skeletal nodes. and This constitutes a complete set of dimensional parameters that control the geometry of the component. The dimensional parameter set is stored in a structured manner, preserving the topological order of the skeleton nodes so that subsequent 3D geometry generation can drive section sweep or loft operations according to the skeleton path order.

[0041] The style parameterization of texture data begins with the gray-level co-occurrence matrix (GLCM) and involves its frequency domain transformation. The GLCM is treated as a two-dimensional discrete signal, and a two-dimensional discrete Fourier transform is applied to obtain its frequency domain representation. In the frequency domain, the frequency components with concentrated energy correspond to the main periodic structure of the texture. The highest-energy dominant frequency components in the spectrum are extracted, and their frequency values ​​are... Texture periodicity parameter reflects the spatial repetition period of the texture. Defined as the reciprocal of the period corresponding to the main frequency component, i.e. The unit is pixel length. The larger the value, the wider the repeating spacing of the texture pattern, corresponding to a sparser texture arrangement; The smaller the size, the denser and more intricate the texture pattern. This is true for periodic patterns commonly found on jewelry surfaces, such as fish scales, woven patterns, and filigree. It can accurately quantify its repeatability scale.

[0042] Phase distribution features are extracted simultaneously in the frequency domain. The phase spectrum corresponding to the dominant frequency component is analyzed, and the phase values ​​are statistically analyzed. The phase distribution pattern within the range. If the phase distribution exhibits a regular linear trend, it indicates that the texture pattern is uniformly translated along a certain direction, corresponding to orthogonal arrangement parameters; if the phase distribution exhibits a random scattering pattern, it indicates that the texture pattern has irregular offsets, corresponding to random scattering arrangement parameters; if the phase distribution exhibits a symmetrical pattern with a certain angle as an axis, it corresponds to mirror symmetry arrangement parameters. These phase distribution characteristics are quantified into texture arrangement parameters. , Use discrete enumeration values ​​to distinguish different permutation types. and These elements are combined to form a set of style parameters that control the appearance of the component's surface. This set of style parameters is also bound to the corresponding component in a structured manner, supporting parameter-driven processing during subsequent procedural texture generation on the 3D model surface.

[0043] When identifying assembly dependencies between connected components based on connection types in the topology connection data, different dependency rules are established for different connection types. For rigid fixed connections, there are no relative degrees of freedom between the two connected components, and their dimensional parameter sets must satisfy geometric fit constraints at the contact area: if the components The cross-sectional width at the contact end is ,part The cross-sectional width at the corresponding contact end is Then the association constraint requires To ensure dimensional consistency at the assembly interface. For movable hinged connections, the two connected components are allowed to rotate about the normal vector direction at the contact point. The associated constraint only restricts the position of the contact point to coincide, not to strictly equalize the cross-sectional dimensions, but requires that the cross-sectional width of the hinged component does not exceed the receiving aperture. ,in The cross-sectional width of the pin component is... This represents the cross-sectional width of the slotted component. For nested enclosure connection types, the inner diameter of the outer component must be greater than or equal to the outer diameter of the inner component to form an enclosure constraint.

[0044] In terms of the correlation constraints between style parameter sets and size parameter sets, connection type also plays a guiding role. For adjacent parts with rigid fixed connections, if the texture arrangement parameters in the contact area are... Inconsistencies trigger style coordination constraints, marking the texture transition in the contact area as a gradient blend to prevent abrupt texture changes on the 3D model surface. For components with active hinged connections, since movement gaps need to be maintained at the hinges, the style parameters include texture periodicity parameters. Minimum constraints must be met near the hinge ends to avoid visual interference from overly dense texture patterns in the moving area. All the aforementioned geometric fit constraints, containment constraints, and style coordination constraints are unified and summarized to form a complete set of associated constraints, which, together with the size parameter set and style parameter set, constitute the parametric representation data. The parametric representation data is organized in a hierarchical structure: the top layer is a list of components, each component node carries its corresponding size parameter set and style parameter set, and the connecting edges between component nodes record the corresponding associated constraints. This comprehensively describes the geometry, surface appearance, and assembly relationships of each component of the ornament, providing a complete parameter-driven foundation for the subsequent automated generation of 3D digital ornaments.

[0045] In one optional embodiment, the assembly dependencies between connected components are identified based on the connection types in the topology connection data. Based on the assembly dependencies, the associated constraints are determined, and the parameterized expression data is obtained, including: When the connection type is embedded connection, extract the cross-sectional dimension parameters of the embedded part and the embedded part in the contact area, and establish a geometric fit constraint that the cross-sectional dimension parameters of the embedded part are smaller than the corresponding cross-sectional dimension parameters of the embedded part. When the connection type is tessellation, extract the texture periodicity parameters of the components on both sides of the tessellation interface, and establish an interface transition constraint that maintains the continuity of the texture periodicity parameters on both sides of the interface. When the connection type is a wrapping connection, extract the length class size parameters of the wrapping component and the wrapped component, and establish a wrapping coverage constraint where the length class size parameter of the wrapping component is greater than the corresponding length class size parameter of the wrapped component. By using geometric fit constraints, interface transition constraints, and wrapping constraints as associated constraints, parameterized representation data is obtained.

[0046] In one specific implementation, after obtaining the size parameter set and style parameter set of each component, it is necessary to identify the assembly dependency relationship between adjacent components based on the connection type in the topology connection data, and establish corresponding association constraints accordingly, ultimately forming a complete parameterized expression data. The assembly dependency relationships corresponding to different connection types have fundamental differences in geometric semantics, so they need to be classified and processed separately, extracting parameters of different dimensions and establishing corresponding constraints.

[0047] For embedded connection types, commonly seen in jewelry structures such as prong settings and channel settings, where a portion of the geometry of one component is embedded into a groove or hole in another component, the cross-sectional dimensional parameters of the embedding and embedded components at the contact area are extracted after identifying the embedded connection. These cross-sectional dimensional parameters include the maximum circumscribed dimension, minimum inscribed dimension, and equivalent diameter of the cross-section, all derived from the dimensional parameter set obtained from contour data conversion. Let the equivalent cross-sectional dimension of the embedding component at the contact end be... The equivalent cross-sectional size of the embedded component in the corresponding contact area is Then the geometric fit constraint requirements This ensures that the embedded component can be physically accommodated within the recessed structure of the embedded component, avoiding interference and collisions. In actual parametric design, this constraint is not limited to the size relationship of a single dimension, but also requires constraints to be applied to each corresponding dimension of the cross-sectional shape. For example, when the cross-section is rectangular, constraints need to be established for the length and width directions respectively; when the cross-section is circular, only the equivalent diameter needs to be constrained. Geometric fit constraints are incorporated as hard constraints into the parametric representation data. During subsequent parameter iteration and adjustment, any... or All modifications must satisfy this inequality relationship; otherwise, a constraint violation warning will be triggered and the system will automatically revert to the valid parameter range.

[0048] For inlaid connections, a common technique in jewelry such as patchwork setting and enamel filling, two components are joined at the interface with a continuous texture transition, requiring visual continuity in the texture patterns on both sides of the interface. After identifying the inlaid connection, the texture periodicity parameters of the components on both sides of the interface are extracted. Texture periodicity parameters. The periodicity parameter of the texture pattern in space, which has been calculated during the conversion between the size parameter set and the style parameter set, is given. Let the periodicity parameter of the texture on one side of the interface be... The texture periodicity parameter of the component on the other side of the interface is Interface transition constraints require that the two maintain continuity at the interface, that is... and The difference between the two textures approaches zero within the allowable error range, or there exists an integer multiple of their frequency, thus ensuring that the texture does not abruptly break at the interface. Specifically, when both textures are periodic and regular, the continuity condition is expressed as the periodic parameters on both sides being equal or in an integer ratio; when one side is a non-periodic texture, the constraint degenerates into the continuity requirement of the texture gradient at the interface, that is, the gray-level transition slope of the textures on both sides of the interface remains consistent. The interface transition constraint is incorporated into the parameterized data as a soft constraint, participating in the optimization in the form of a penalty term during parameter iteration, guiding the texture parameters to adjust in a direction that satisfies continuity, rather than forcibly truncating parameter modification operations.

[0049] For the wrap-around connection type, commonly seen in jewelry as gold-rimmed edges or metal encasing the base of gemstones, a component extends along the contour of the component it wraps around, partially or completely covering it. After identifying the wrap-around connection, the length-type dimensional parameters of the wrapping and wrapped components are extracted. These length-type dimensional parameters include the extension length along the main axis, the circumferential coverage arc length, and the coverage height, etc., which are also derived from the dimensional parameter set obtained by converting the contour data. Let the length-type dimensional parameters of the wrapping component along the coverage direction be... The length-like dimension parameter of the wrapped part in the corresponding direction is: Then the package coverage constraint requirement This ensures that the wrapping component geometrically and completely covers the target area of ​​the wrapped component without any gaps. When the wrapping is designed for partial coverage, the constraint is adjusted so that the length of the wrapping component is greater than the product of a preset coverage ratio and the length of the wrapped component. The coverage ratio is specified by the user or automatically identified from the image data as a design intent parameter. The wrapping coverage constraint is also incorporated as a hard constraint into the parametric representation data. During the parameter-driven geometry generation stage, if the length of the wrapping component does not meet the coverage condition, it will be automatically... Stretch to the minimum valid value that satisfies the constraints.

[0050] The three types of constraints mentioned above—geometric fit constraints, interface transition constraints, and wrapping / covering constraints—are uniformly stored as associated constraints in the parametric representation data. Structurally, the parametric representation data comprises three subsets: a set of size parameters, a set of style parameters, and a set of associated constraints. The set of associated constraints is organized in the form of a directed constraint graph, where each directed edge represents a constraint relationship. The start and end points of the edge correspond to the constrained parameter nodes, and the edges are accompanied by constraint type identifiers and constraint expressions. This structured storage method enables iterative adjustments to the geometry of each component in a parametrically driven manner during the subsequent 3D digital jewelry generation stage. Each parameter modification automatically triggers the verification and propagation of associated constraints, ensuring the geometric rationality and technological feasibility of the overall jewelry structure.

[0051] When dealing with jewelry structures featuring complex connections, the same pair of adjacent components may involve combinations of multiple connection types. For example, the base of a gemstone and its support may have both embedded and wrapping connections. In this case, geometric fit constraints and wrapping constraints are established for the pair of components respectively. These two types of constraints are stored side-by-side in the associated constraint set and are simultaneously verified during parameter iteration. The parameters are corrected based on the most stringent constraint conditions, thereby ensuring that the parametric design results under multi-constraint scenarios satisfy all assembly dependencies.

[0052] The generation of the geometry of each component is driven by the set of dimensional parameters in the parameterized expression data, the spatial arrangement of each component is controlled by the assembly angle in the topological connection data, and the dimensional parameter set is iteratively adjusted by associative constraints to generate a three-dimensional digital ornament, including: Construct the center curves of each component based on the skeleton nodes in the size parameter set, calculate the curvature tensor of the center curve at each skeleton node, determine the placement posture of the cross-sectional profile based on the principal direction of the curvature tensor, perform non-uniform scaling in the principal direction based on the cross-sectional size parameters, and generate the solid geometry of each component by sweeping the deformed cross-sectional profile along the center curve. Boundary curves of the contact area are extracted from the topological connection data. Surface patches matching the boundary curves are identified on the solid geometry surfaces of each component. The normal vector field distribution of the surface patches is calculated. The assembly reference direction is obtained by weighted averaging of the normal vector field. The assembly angle is decomposed into a rotation component around the assembly reference direction and a translation component perpendicular to the assembly reference direction. Rotation transformation and translation transformation are performed in sequence to realize the spatial configuration of each component. After spatial configuration, a distance field is established between connected components based on the geometric fit constraints in the associated constraints. The adjustment direction of the size parameter set is determined by solving the gradient direction of the distance field. The adjustment range of the size parameter set is determined based on the magnitude of the distance field. The size parameter set is updated and the solid geometry is regenerated. This process is iterated until the distance field converges, generating a three-dimensional digital ornament.

[0053] After obtaining the parametric representation data and topological connection data of each component, the abstract set of parameters needs to be transformed into a three-dimensional entity with a real geometric shape, and the precise assembly between the components in space needs to be completed. This process begins with the construction of the center curve and ends with iterative constraint convergence, spanning three tightly coupled stages: geometry generation, spatial configuration, and adaptive parameter adjustment.

[0054] The dimensional parameter set contains a sequence of skeleton nodes for each component. Each skeleton node records its coordinates in 3D space and the corresponding cross-sectional dimension parameters. The skeleton node sequences are connected sequentially in topological order, and a smooth center curve is fitted using cubic spline interpolation, ensuring that the curve satisfies the following conditions at each node. Continuity conditions are established to ensure the smoothness of the cross-sectional transition during subsequent sweeping operations. After the center curve is generated, the curvature tensor at each skeleton node on the curve is calculated. Let the unit tangent vector of the curve at the node be... The curvature vector is Then the curvature tensor Constructed by the cross product relationship between the tangent vector and the curvature vector, its principal direction corresponds to the normal of the bending plane with the maximum curvature. Based on the principal direction of the curvature tensor, the placement posture of the cross-sectional profile at this node is determined, that is, the plane containing the cross-section is perpendicular to the center curve, and the principal axis of the cross-section is aligned with the principal direction of curvature, so that the cross-sectional profile can rotate naturally with the curve and avoid torsional distortion of the cross-section.

[0055] The shape of the cross-sectional profile is described by the cross-sectional class parameters in the dimension parameter set, including the cross-sectional width. (corresponding to the first) The local cross-sectional width at each skeleton node (consistent with the meaning in claim 4) and cross-sectional height parameters After determining the placement orientation, non-uniform scaling is applied to the cross-sectional profile along the principal and secondary directions of curvature, with the scaling ratio determined by... and The ratio relationship determines the gradient shape of the cross-section at different nodes, which corresponds to the original ornament image. After setting the orientation and non-uniform scaling of the cross-section at each node, a sweep operation is performed on the deformed cross-sectional contour sequence along the center curve. That is, using the center curve as a guide, the transition area between adjacent cross-sections is filled by linear interpolation or Hermite blending, ultimately generating a closed solid geometry. For branched structures or multi-segment spliced ​​parts, after performing the above sweep process on each segment, the solid segments are merged into a single part geometry through a Boolean merge operation.

[0056] After the solid geometry of each component is generated, the components need to be correctly positioned in space based on the assembly angles in the topology connection data. The boundary curve of the contact area is extracted from the topology connection data; this boundary curve describes the contour boundary of two adjacent components at the contact position. On the surface of the solid geometry of each component, a surface sampling and curve matching algorithm is used to identify the surface patch that best matches the geometric shape of the contact area boundary curve. Specifically, the surface of the solid geometry is uniformly and discretely sampled, and the surface normal vector at each sampling point is calculated to form a normal vector field distribution. Then, within the candidate surface patch region, the normal vectors of all sampling points are weighted and averaged. The weight is determined by the reciprocal of the distance from the sampling point to the boundary curve; the closer the sampling point is, the greater its contribution to the assembly reference direction. Finally, the assembly reference direction of the contact area is obtained. .

[0057] Assembly angle (Consistent with the meaning in claim 3, indicating that the first...) The rotational transformation matrix corresponding to each contact cluster is decomposed into two independent transformation components at this stage: around the assembly reference direction. rotational component Translational component perpendicular to the assembly reference direction During the decomposition process, firstly... The corresponding rotation axis is projected onto the assembly datum direction, the rotation angle around that direction is extracted, and the rotation component is constructed. The remaining positional offset is then decomposed into a translation vector perpendicular to the assembly reference direction. The component geometry is subjected to rotation and translation transformations sequentially. The rotation transformation is centered on the centroid of the contact area, and the translation transformation is performed along... The direction is moved by a corresponding distance, thereby precisely aligning each component to the spatial configuration required by the design intent.

[0058] After initial spatial configuration, minor geometric mismatches may exist between adjacent components in the contact area, manifesting as local gaps or slight interference. To eliminate these mismatches, an iterative adjustment mechanism based on association constraints is introduced. The geometric fit constraints within the association constraints describe the geometric relationships that connected components should satisfy in the contact area, such as contact surface fit and cross-sectional width coordination. In each iteration, a directed distance field is established between adjacent components. , The value at each point in space represents the signed distance from that point to the nearest contact interface; a positive value indicates that it is located outside the component, and a negative value indicates that it is located inside the component. This is achieved through calculation... gradient field Determine the adjustment direction of each parameter in the dimensional parameter set: If a component has a gap at the contact end ( If interference exists, then the corresponding cross-sectional width or length parameter increases along the gradient direction; If the gradient is zero, then the corresponding parameter decreases in the opposite direction of the gradient. The adjustment magnitude is determined proportionally to the magnitude of the distance field; let the adjustment step size coefficient be... Then the parameter update amount and It is proportional to ensure that the adjustment range is larger when the gap or interference is large, and the adjustment range tends to be stable when it is close to convergence.

[0059] After each parameter update, the center curve construction, section sweep, and spatial configuration processes are re-executed to generate a new solid geometry, and the distance field is recalculated. The iteration terminates when the maximum absolute value of the distance field across all contact regions falls below a preset convergence threshold. This means that the geometric errors at the contact interface of all adjacent components meet the accuracy requirements. The resulting 3D digital ornament is highly consistent with the original ornament image in terms of geometry, the assembly relationship between the components is accurate, and the size parameter set and style parameter set have been adaptively optimized, which can be directly used for subsequent parametric design output and process data generation.

[0060] In one optional embodiment, generating the solid geometry of each component by sweeping the deformed cross-sectional profile along the center curve includes: The center curve is parameterized by arc length to obtain a parameterized center curve. A sweep sampling point sequence is obtained by uniformly sampling at arc length intervals on the parameterized center curve. The tangent vector of the parameterized center curve at each sweep sampling point in the sweep sampling point sequence is calculated. The tangent vector is used as the sweep direction of the cross section profile. A cross section placement plane perpendicular to the tangent vector is established at each sweep sampling point. The deformed cross section profile is projected onto the cross section placement plane to obtain the cross section instance at each sweep sampling point. Calculate the torsion angle of the parameterized center curve between adjacent sweep sampling points in the sweep sampling point sequence. When the torsion angle exceeds the preset torsion threshold, insert a compensation sampling point between adjacent sweep sampling points. Calculate the torsion angle difference between the compensation sampling point and the adjacent sweep sampling point. Perform rotation compensation on the cross-sectional instance at the compensation sampling point based on the torsion angle difference. Extract the contour control points of each cross-section instance, construct a connecting surface between the corresponding contour control points of adjacent cross-section instances, obtain surface mesh patches by triangulating the connecting surface, topologically stitch all surface mesh patches to form a closed surface, fill the closed surface with voxels, and generate the solid geometry of each component.

[0061] In one specific implementation, when parameterizing the center curve, the total arc length of the center curve is normalized to a unit interval, so that the parameter difference between any two points on the parameterized curve is proportional to the corresponding arc length. Specifically, the arc length is accumulated segment by segment along the center curve, and the accumulated arc length at each control point is divided by the total arc length to obtain the normalized parameter value corresponding to each control point, thereby establishing the parameterized center curve. Uniform sampling is performed on the parameterized center curve at arc length intervals. The sampling interval is adaptively determined based on the overall size and cross-sectional complexity of the component. In areas with drastic cross-sectional changes, the sampling interval can be appropriately reduced to improve geometric reconstruction accuracy. After uniform sampling, a sweep sampling point sequence is obtained. The arc length distribution of each sampling point on the curve is uniform, ensuring that the subsequently generated solid geometry has a consistent mesh resolution along the curve direction.

[0062] At each sweep sampling point, the tangent vector at that point is obtained by taking the first derivative of the parameterized center curve and normalizing it. The tangent vector represents the extension direction of the curve at that point. Using the tangent vector as the sweep direction of the cross-sectional profile, a cross-sectional placement plane perpendicular to the tangent vector is established at each sweep sampling point. The normal vector of the cross-sectional placement plane is the tangent vector at that point, and the plane passes through the sweep sampling point itself. The deformed cross-sectional profile is projected onto the cross-sectional placement plane. During the projection process, the relative shape of the cross-sectional profile in the plane remains unchanged, and only the spatial orientation of the cross-sectional profile is adjusted to align it with the cross-sectional placement plane, thus obtaining the cross-sectional instances at each sweep sampling point. The dimensional parameters of the cross-sectional instances are derived from the dimensional parameter set in the parameterized representation data, ensuring that the width and height of the cross-section at each sampling point continuously interpolate with the local dimensional parameters at the skeleton nodes.

[0063] After obtaining the sweep sampling point sequence, the torsion angle of the parameterized center curve between adjacent sweep sampling points is calculated. The torsion angle reflects the amount of rotation of the normal vector of the cross-section placement plane around the tangent vector during propagation along the curve, and is usually obtained by the torsion integral of the curve. When the torsion angle between adjacent sampling points exceeds a preset torsion threshold, without compensation, there will be a significant torsional jump between adjacent cross-section instances, resulting in distortion of the generated solid geometry surface. To eliminate this problem, compensation sampling points are inserted between adjacent sweep sampling points where the torsion angle exceeds the threshold. The positions of the compensation sampling points are inserted according to the arc length, and the number is determined by the ratio of the torsion amount exceeding the threshold to the threshold, so that the torsion angle between any adjacent sampling points does not exceed the preset torsion threshold. The difference in torsion angle between the compensation sampling point and its adjacent sweep sampling points is calculated. Based on this difference, the cross-section instance at the compensation sampling point is rotated for compensation, with the rotation axis being the tangent vector at that point, and the rotation amount equal to the corresponding torsion angle difference, thereby making the cross-section instance transition smoothly along the curve direction and eliminating the geometric distortion caused by the torsion jump.

[0064] When extracting the contour control points of each cross-section instance, sampling is performed uniformly along the cross-section contour according to the arc length to extract a sequence of control points at equal intervals. For cross-section contours with characteristic corners or sharp turns, additional control points are inserted at the characteristic locations to preserve the geometric features of the cross-section. A one-to-one correspondence is established between adjacent cross-section instances according to the sequential index of the control points, and a connecting surface is constructed between the corresponding control points. The connecting surface is generated using bilinear or bicubic interpolation to ensure sufficient smoothness in both the cross-section direction and the sweep direction. When the number of control points in adjacent cross-section instances is inconsistent, resampling is used to unify the number of control points to the one with the larger number before establishing the correspondence, thus avoiding topological errors on the surface.

[0065] When triangulating the connected surfaces, each quadrilateral mesh cell of the connected surface is divided into two triangles along its diagonal, resulting in surface mesh patches. During triangulation, diagonal directions that make the aspect ratio of the triangles close to 1 are preferred to improve mesh quality and reduce geometric errors in subsequent voxelization. All surface mesh patches are topologically stitched together, that is, the shared edges of adjacent surface mesh patches are merged to eliminate duplicate vertices and boundary gaps, forming closed surfaces. The topological integrity of the closed surfaces is verified using Euler's formula to ensure that the surfaces satisfy the topological conditions of closed manifolds. For the open sections at the ends of the components, end cap meshes are generated using Delaunay triangulation with the contour control points of the end section instances as boundaries. The end cap meshes are then stitched together with the side meshes to complete the construction of the complete closed surface.

[0066] When performing voxelization filling on a closed surface, the bounding box of the component is used as the voxelization region, and the bounding box is divided into a uniform voxel mesh according to a preset voxel resolution. For each voxel within the bounding box, it is determined whether its center point is located inside the closed surface. The determination method uses ray casting, that is, a ray is emitted from the center point of the voxel along a fixed direction, and the number of intersections between the ray and the closed surface is counted. If the number of intersections is odd, the voxel is determined to be inside the surface. All internal voxels are marked as solids, and external voxels are marked as empty, thus obtaining the voxelized solid representation of the component. The voxel resolution is determined based on the minimum feature size of the component, ensuring that the minimum feature is covered by at least 3 voxels to guarantee the complete restoration of geometric details. Finally, the voxelization result is converted into a polyhedral solid geometry. The Marching Cubes algorithm is used to extract isosurfaces from the voxel mesh to obtain smooth solid geometry of each component, providing a complete geometric foundation for the subsequent assembly and parametric design of 3D digital ornaments.

[0067] A second aspect of this invention provides a parametric recognition and design system for jewelry structures that integrates computer vision, comprising: An image acquisition unit is used to acquire image data of the jewelry to be identified. The semantic segmentation unit is used to perform semantic segmentation on image data, identify each component and its boundary in the jewelry to be identified, and obtain the contour data and texture data of each component. The topology extraction unit is used to extract the set of contact points and contact surface normals between components based on the boundary positional relationship of each component, determine the connection type and assembly angle, and obtain topology connection data. The parameter transformation unit is used to convert contour data into a size parameter set and texture data into a style parameter set. It establishes the association constraints between the size parameter set and the style parameter set according to the connection type in the topological connection data to obtain parameterized expression data. The three-dimensional generation unit is used to drive the geometric generation of each component with the set of size parameters in the parameterized expression data, control the spatial arrangement of each component with the assembly angle in the topological connection data, and iteratively adjust the set of size parameters with associated constraints to generate a three-dimensional digital ornament. The result output unit is used to carry the parameterized expression data and the topological connection data of the three-dimensional digital ornament, and output the three-dimensional digital ornament to realize the parameterized design of the ornament.

[0068] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0069] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0070] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A parametric recognition and design method for jewelry structures integrating computer vision, characterized in that, include: Acquire image data of the jewelry to be identified; Semantic segmentation is performed on the image data to identify the components and boundaries of the jewelry to be identified, and the contour data and texture data of each component are obtained. Based on the boundary positional relationship of each component, the set of contact points and contact surface normal vectors between the components are extracted to determine the connection type and assembly angle, thus obtaining topological connection data; Contour data is converted into a set of size parameters, texture data is converted into a set of style parameters, and association constraints between the set of size parameters and the set of style parameters are established based on the connection type in the topological connection data to obtain parameterized representation data. The geometry of each component is generated by using the set of size parameters in the parameterized expression data, the spatial arrangement of each component is controlled by the assembly angle in the topological connection data, and the set of size parameters is iteratively adjusted by the associated constraints to generate a three-dimensional digital ornament. The three-dimensional digital ornament carries the parametric expression data and the topological connection data, and outputs the three-dimensional digital ornament to realize the parametric design of the ornament.

2. The method according to claim 1, characterized in that, Semantic segmentation is performed on the image data to identify the components and boundaries of the jewelry to be identified, resulting in contour and texture data of each component, including: Multi-scale feature extraction is performed on image data. Spatial structure features and color distribution features of the image are extracted at different scale levels. The spatial structure features and color distribution features at each scale level are integrated through feature fusion operation to obtain multi-scale fused features. A pixel-level semantic prediction map is constructed based on multi-scale fusion features. A region growing operation is performed on the semantic prediction map to aggregate pixels with the same semantic category and continuous spatial location into candidate component regions. The region boundaries are extracted for each candidate component region, and the pixel coordinate sequence on the region boundary is used as the initial expression of the boundary contour. Curve fitting is performed on the initial representation of the boundary contour. By calculating the curvature change of the fitted curve, corner features and curve segment features on the boundary are identified. The corner features and curve segment features are organized into a segmented contour representation, which serves as the contour data of each component. Texture feature statistics are performed on the pixels in each candidate component region, the gray-level co-occurrence matrix of the pixels in the region is calculated, and texture directionality index and texture roughness index are extracted from the gray-level co-occurrence matrix. The texture directionality index and texture roughness index are used as the texture data of each component.

3. The method according to claim 1, characterized in that, Based on the boundary positional relationships of each component, the set of contact points and the normal vector of the contact surface between the components are extracted to determine the connection type and assembly angle, resulting in topological connection data including: A directed distance field is constructed from the contour data of each component. The boundary normal vector distribution is obtained by solving the gradient field of the directed distance field. The directed distance fields of adjacent components are superimposed, and the set of contact points is extracted from the zero iso-surface after the directed distance fields are superimposed. A topological connectivity analysis is performed on the set of contact points, and the topologically connected contact points are divided into contact clusters. A local coordinate system is constructed for each contact cluster, and the variance distribution of the contact cluster along each coordinate axis in the local coordinate system is calculated by principal component analysis to obtain the morphological tensor of the contact cluster. Eigenvalue decomposition is performed on the morphological tensor to obtain the morphological principal axis directions and the corresponding morphological contribution. The connection type is determined according to the distribution pattern of the morphological contribution. When the morphological contribution is concentrated in a single morphological principal axis direction, it is determined to be an embedded connection. When the morphological contribution is distributed in two orthogonal morphological principal axis directions, it is determined to be an inlaid connection. When the morphological contribution is evenly distributed in each morphological principal axis direction, it is determined to be a wrapping connection. The boundary normal vector distribution of each contact cluster is spherically fitted, and the extreme direction of the fitted sphere is extracted as the contact surface normal vector. The assembly angle is obtained by calculating the rotation transformation matrix between the contact surface normal vector and the geometric center of the component. Topological connection data containing connection type, contact point set, contact surface normal vector and assembly angle is constructed.

4. The method according to claim 1, characterized in that, The contour data is converted into a size parameter set, and the texture data is converted into a style parameter set. Association constraints between the size parameter set and the style parameter set are established based on the connection types in the topological connectivity data, resulting in parameterized representation data including: Skeleton extraction is performed on the contour data. The skeleton structure of the component is obtained by calculating the central axis transformation of the contour. Skeleton nodes and skeleton branches are marked on the skeleton structure. The geodesic distance between skeleton nodes is used as a length-type dimension parameter, and the local contour width at the skeleton node is used as a cross-section-type dimension parameter, thus forming a set of dimension parameters that control the shape of the component. The gray-level co-occurrence matrix in the texture data is transformed in the frequency domain to extract the dominant frequency component and phase distribution features of the texture. By mapping the dominant frequency component to the texture periodicity parameter and the phase distribution feature to the texture arrangement parameter, a set of style parameters for controlling the appearance of the component is formed. Based on the connection type in the topology connection data, the assembly dependency relationship between connected components is identified. Based on the assembly dependency relationship, the association constraint is determined, and the parameterized expression data is obtained.

5. The method according to claim 4, characterized in that, Based on the connection types in the topology connection data, the assembly dependencies between connected components are identified. Based on the assembly dependencies, the associated constraints are determined, and the parameterized representation data is obtained, including: When the connection type is embedded connection, extract the cross-sectional dimension parameters of the embedded part and the embedded part in the contact area, and establish a geometric fit constraint that the cross-sectional dimension parameters of the embedded part are smaller than the corresponding cross-sectional dimension parameters of the embedded part. When the connection type is tessellation, extract the texture periodicity parameters of the components on both sides of the tessellation interface, and establish an interface transition constraint that maintains the continuity of the texture periodicity parameters on both sides of the interface. When the connection type is a wrapping connection, extract the length class size parameters of the wrapping component and the wrapped component, and establish a wrapping coverage constraint where the length class size parameter of the wrapping component is greater than the corresponding length class size parameter of the wrapped component. By using geometric fit constraints, interface transition constraints, and wrapping constraints as associated constraints, parameterized representation data is obtained.

6. The method according to claim 1, characterized in that, The generation of the geometry of each component is driven by the set of dimensional parameters in the parameterized expression data, the spatial arrangement of each component is controlled by the assembly angle in the topological connection data, and the dimensional parameter set is iteratively adjusted by associative constraints to generate a three-dimensional digital ornament, including: Construct the center curves of each component based on the skeleton nodes in the size parameter set, calculate the curvature tensor of the center curve at each skeleton node, determine the placement posture of the cross-sectional profile based on the principal direction of the curvature tensor, perform non-uniform scaling in the principal direction based on the cross-sectional size parameters, and generate the solid geometry of each component by sweeping the deformed cross-sectional profile along the center curve. Boundary curves of the contact area are extracted from the topological connection data. Surface patches matching the boundary curves are identified on the solid geometry surfaces of each component. The normal vector field distribution of the surface patches is calculated. The assembly reference direction is obtained by weighted averaging of the normal vector field. The assembly angle is decomposed into a rotation component around the assembly reference direction and a translation component perpendicular to the assembly reference direction. Rotation transformation and translation transformation are performed in sequence to realize the spatial configuration of each component. After spatial configuration, a distance field is established between connected components based on the geometric fit constraints in the associated constraints. The adjustment direction of the size parameter set is determined by solving the gradient direction of the distance field. The adjustment range of the size parameter set is determined based on the magnitude of the distance field. The size parameter set is updated and the solid geometry is regenerated. This process is iterated until the distance field converges, generating a three-dimensional digital ornament.

7. The method according to claim 6, characterized in that, The solid geometry of each component is generated by sweeping the deformed cross-sectional profile along the center curve, including: The center curve is parameterized by arc length to obtain a parameterized center curve. A sweep sampling point sequence is obtained by uniformly sampling at arc length intervals on the parameterized center curve. The tangent vector of the parameterized center curve at each sweep sampling point in the sweep sampling point sequence is calculated. The tangent vector is used as the sweep direction of the cross section profile. A cross section placement plane perpendicular to the tangent vector is established at each sweep sampling point. The deformed cross section profile is projected onto the cross section placement plane to obtain the cross section instance at each sweep sampling point. Calculate the torsion angle of the parameterized center curve between adjacent sweep sampling points in the sweep sampling point sequence. When the torsion angle exceeds the preset torsion threshold, insert a compensation sampling point between adjacent sweep sampling points. Calculate the torsion angle difference between the compensation sampling point and the adjacent sweep sampling point. Perform rotation compensation on the cross-sectional instance at the compensation sampling point based on the torsion angle difference. Extract the contour control points of each cross-section instance, construct a connecting surface between the corresponding contour control points of adjacent cross-section instances, obtain surface mesh patches by triangulating the connecting surface, topologically stitch all surface mesh patches to form a closed surface, fill the closed surface with voxels, and generate the solid geometry of each component.

8. A parametric recognition and design system for jewelry structures integrating computer vision, used to implement the method as described in any one of claims 1-7, characterized in that, include: An image acquisition unit is used to acquire image data of the jewelry to be identified. The semantic segmentation unit is used to perform semantic segmentation on image data, identify each component and its boundary in the jewelry to be identified, and obtain the contour data and texture data of each component. The topology extraction unit is used to extract the set of contact points and contact surface normals between components based on the boundary positional relationship of each component, determine the connection type and assembly angle, and obtain topology connection data. The parameter transformation unit is used to convert contour data into a size parameter set and texture data into a style parameter set. It establishes the association constraints between the size parameter set and the style parameter set according to the connection type in the topological connection data to obtain parameterized expression data. The three-dimensional generation unit is used to drive the geometric generation of each component with the set of size parameters in the parameterized expression data, control the spatial arrangement of each component with the assembly angle in the topological connection data, and iteratively adjust the set of size parameters with associated constraints to generate a three-dimensional digital ornament. The result output unit is used to carry the parameterized expression data and the topological connection data of the three-dimensional digital ornament, and output the three-dimensional digital ornament to realize the parameterized design of the ornament.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images